Validation of cardiovascular risk prediction in chinese adults (P-CARDIAC) with territory-wide data

Abstract Introduction Risk prediction tools for early identification and proactive prevention of cardiovascular (CV) events. Personalized CARdiovascular DIsease risk Assessment for Chinese (P-CARDIAC) is a machine-learning driven model developed based on Hong Kong population. It estimates the 10 years of incident (primary model) and recurrent (secondary model) CV risk for individuals. The model considered an array of risk variables such as diagnoses, prescriptions, procedures, laboratory tests, and healthcare service utilization in structured data format from electronic health records(EHR). The performance of P-CARDIAC is shown superior to TRS-2°P and SMART2 in previous study 1. Purpose We aim to validate P-CARDIAC primary and secondary models with territory-wide data from Hong Kong. Methods Territory-wide longitudinal records were obtained from the Hospital Authority. All participants who have ever received a lipid test from 2004 to 2021 at any service points managed by the Hospital Authority were identified. Participants with no prior CV events defined by ICD-9 codes (such as: peripheral artery diseases, coronary heart diseases, myocardial infarction, stroke and revascularization) were eligible for primary model validation, whilst those with prior CV events were eligible for secondary model validation. We validated the model performance with the comparison of common risk scores, including PCE, PREDICT, China-PAR, Framingham (Asian), TRS-2°P and SMART2 using the validation cohorts with 1000 bootstrap replicates. Results For primary model, 2,638,820 participants without any prior CV events were identified. For secondary model, 549,650 participants with prior CV events were identified. P-CARDIAC primary model showed good C-statistics of 0.84 whilst other similar risk prediction scores showed less optimal performance C-statistics: PCE (white) 0.66; PCE (African) 0.63; PREDICT 0.67; China-PAR 0.65, Framingham (Asian) 0.65. Similarly, P-CARDIAC secondary model showed good C-stats of 0.65 than other risk scores C-stats: SMART2 0.51; TRS-2°P 0.50. Both P-CARDIAC models also showed best calibration-in-the-large compared with other risk scores. Conclusion P-CARDIAC models performed the best in validation. Risk prediction model developed with local data performs better than existing risk scores. Given that healthcare resources are finite, locally developed tools can help identify people at higher risk of CV events and offer appropriate early intervention.